Long-running Agents

Abstract

The transition from ephemeral, chat-based AI interactions to long-running agents necessitates a shift toward persistent architectures that maintain progress across multiple sessions and context windows. To achieve multi-day task execution, agents must decouple reasoning logic from ephemeral execution sandboxes and durable event logs. This architectural separation allows for robust state management, enabling agents to recover from process failures and resume complex tasks without context loss. Essential design patterns for these systems include the functional separation of planners, workers, and judges to overcome self-verification biases and prevent premature task completion. Implementation strategies—such as checkpoint-and-resume mechanisms, persistent memory layering, and structured session handoffs—address the technical limitations of finite context windows and performance degradation over time. While current industrial trends favor managed runtimes that productize these primitives, significant challenges remain concerning operational costs, the security of autonomous tool use, and alignment drift during extended reasoning horizons. Success in deploying long-running agents depends on the engineering of external state layers and rigorous verification protocols that ensure continuous progress toward defined, testable completion criteria. – AI-generated abstract.